Flow Matching for Molecules (FM3)

FM3 learns an Optimal-Transport vector field that pushes pure Gaussian noise into realistic 1D molecular embeddings (SELFIES/SMILES). The field is parameterized by a Mamba-inspired multi-scale CNN that regresses the instantaneous velocity vθ(xt,t).

For the 2.1M curated and filtered data from ChemBL34, COCONUTDB, and SuperNatural3, you can e-mail me [email protected]

File	What it does
fm3.py	full model + schedulers + backbones
backbones.vectorf	Mamba-style selective mixing + dilated gated CNNs
backbones.simple	2-layer CNN baseline (fast, interpretable)
OptimalTransportSchedule	closed-form α(t)=t, σ(t)=1–t, dα=1, dσ=–1
TimeAdditiveEmbedder	sinusoidal → MLP → residual (no param explosion)

Early Test Result on Small Data (24K)

trained for 10 'epochs' (steps = 10 * len(dataloader)) using discrete trainer and simple backbone:

NUM_SAMPLES = 6
MAX_LEN = 12
STEPS = 200
MODE = "discrete"  # choose: "continuous" or "discrete"
TEMPERATURE = 0.35

References

@misc{lipman2024flowmatchingguidecode,
      title={Flow Matching Guide and Code}, 
      author={Yaron Lipman and Marton Havasi and Peter Holderrieth and Neta Shaul and Matt Le and Brian Karrer and Ricky T. Q. Chen and David Lopez-Paz and Heli Ben-Hamu and Itai Gat},
      year={2024},
      eprint={2412.06264},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2412.06264}, 
}

@misc{lipman2023flowmatchinggenerativemodeling,
      title={Flow Matching for Generative Modeling}, 
      author={Yaron Lipman and Ricky T. Q. Chen and Heli Ben-Hamu and Maximilian Nickel and Matt Le},
      year={2023},
      eprint={2210.02747},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2210.02747}, 
}